Soulempty / Unsupervised-Sketch-to-Photo-Synthesis

Unsupervised-Sketch-to-Photo-Synthesis

This is the official released code for our paper, Unsupervised Sketch to Photo Synthesis, which has been accepted as an Oral paper of ECCV 2020. Please find further details in the project page.

@article{liu2020unsupervised,
  author     = {Runtao Liu and
                Qian Yu and
                Stella Yu},
  title      = {Unsupervised Sketch to Photo Synthesis},
  conference = {ECCV 2020},
  year       = {2020}
}

Environment:

Please install Pytorch 1.4.0 and other requirements like visdom.

Dataset and Released Model

We prepared an example dataset of the shoes and it can be downloaded here.

We also released our trained model checkpoint here.

Testing with Released Model

Testing command for step 1:

python -u test.py \
    --only_fakephoto \
    --sp_attention 1 \
    --load_size 128 \
    --crop_size 128 \
    --results_dir ./results/step1_release_model/ \
    --input_nc 1 \
    --output_nc 1 \
    --epoch 480 \
    --dataroot ../datasets/shoes \
    --name shoes \
    --checkpoints_dir ../experiments/step1/step1_release_model/model/

After the testing of step 1, the output results should be placed at the path datasets/shoes_step2/testA.

Testing command for step 2:

python -u test.py \
    --only_fakephoto \
    --netG resnet_12blocks \
    --dataroot ../datasets/shoes_step2 \
    --load_size 128 \
    --crop_size 128 \
    --results_dir ./results/step2_release_model/ \
    --input_nc 3 \
    --output_nc 3 \
    --epoch 290 \
    --name shoes \
    --checkpoints_dir ../experiments/step2/step2_release_model/model/

We has updated the checkpoint whose performance is slightly better than that reported in the paper. The FID values on the shoe dataset of released models are:

step1: 44.35

step2: 48.20 (report: 48.73)

Note that the FID of step1 should be calculated between the output and grayscale images with 128x128 resolution. The FID of step2, which is the final FID result, should be calculated between the output and RGB images with 128x128 resolution.

Training

Step 1 training command:

python train.py \
    --batch_size 1 \
    --sp_attention 1 \
    --display_freq 20 \
    --prob_add_noise 0.2 \
    --prob_add_patch 0.3 \
    --load_size 140 \
    --crop_size 128 \
    --input_nc 1 \
    --output_nc 1 \
    --niter 100 \
    --niter_decay 600 \
    --dataroot ../datasets/shoes \
    --name shoes \
    --display_port 30001 \
    --checkpoints_dir ./checkpoints/

Step 2 training command:

python train.py \
    --netG resnet_12blocks \
    --loss_style_weight 2 \
    --loss_content_weight 1 \
    --pcploss_weight 0.05 \
    --L1loss_weight 10.0 \
    --display_freq 40 \
    --print_freq 100 \
    --load_size 140 \
    --crop_size 128 \
    --niter 100 \
    --niter_decay 600 \
    --input_nc 3 \
    --output_nc 3 \
    --dataroot ../datasets/shoes_step2 \
    --name shoes \
    --lambda_B 10.0 \
    --display_port 30000 \
    --batch_size 1 \
    --checkpoints_dir ./checkpoints/